Canon EOS 1DX announced - March 2012 availability

[BJL]

would love a reference. do you have one handy?

just so I understand... wouldn't a smaller pixel capture fewer photons? and therefore transmit fewer pieces of information?

Hopefully, other more technically expert members can provide references, but let me point out that you just made the same mistake I was talking about: comparing at equal pixel count (in this case, pixel count of one: the maximum electron capacity and photon count of a single pixel), not allowing for the benefits of having more of those smaller pixels gathering the light from the same part of the image. An over simplified example: halving the photosites area will give about half the election capacity per photosite, but will also give twice as many photosites detecting light from the same part of the image, so the "maximum photo-electron count" for any given part of the image is about the same; it is just divided up as the sum or more, smaller numbers, and this retains more information about where the light comes from.

Pushing pixel downsizing too far will cause degradation, with election capacity per unit area declining when photosite size gets too close to the minimum feature size of the semiconductor fabrication technology, and this might be an issue with compact camera sensors with pixel sizes approaching one micron, but DSLR sensors are nowhere near that limit. In fact they are usually made with older fabrication equipment of far larger minimum feature size than newer equipment, because DSLR photosites are so large compared to what current technology can make.

P.S. 18MP or even less is plenty for my purposes (mostly wildlife and nature close-ups, plus casually documenting life around me), but for many landscape and archtecural  photographers who can often work at minimum ISO speed, 18MP is short of what MF film offered, so increasing resolution beyond that by a factor of two say would be a gain that easily outweighs excessive concern about noise seen only in unnaturally close scrutiny of deep shadows and such.

[ejmartin]

would love a reference. do you have one handy?

just so I understand... wouldn't a smaller pixel capture fewer photons? and therefore transmit fewer pieces of information?

There are two dominant sources of noise in typical images -- photon noise and read noise.  If we're only concerned with photon noise (ie there are enough photons that read noise is a secondary consideration), then smaller pixels are more 'informative'.  The number of photons captured is the same, they are simply partitioned into samples of finer spatial resolution, so the total captured information is larger due to the better spatial localization of any given captured photon.  Read noise, on the other hand, is a constant noise per pixel, so unless the smaller pixels have smaller read noise in proportion to the pixel spacing, they degrade the image by increasing the noise relative to the number of photons captured.

As usual, there are caveats.  If pixels get too small, there can be cross-talk -- red photons leak into green pixels etc, resulting in a loss of color separation and resolution.  Also, lower S/N per pixel leads to a greater degree of interpolation error in the demosaic process, and thus more 'digital artifacting', though that artifacting is pushed off to finer spatial scales where it may be less noticeable.

[BernardLanguillier]

Yes, you averred in a previous post that a company could put > 30 Mp on a 24 x 36 chip. But they would be smaller, and have a smaller photon "basket" if current technology was simply scaled.

Does N or C have a 30+ Mp design in a 24 x 36 chip that actually is "better" from a image and noise perspective? I mean, yes they can do it, but if it brings more issues, why bother?

My understanding is that there is no difference in the behavior of a sensor whether it is APS sized or FF sized.

So you can analyze the behavior of APS sensors on the market today, the better ones using Sony sensors, to extrapolate the behavior of pixels in a 30+ MP sensor.

Measurements do show that the current 16MP sensors on APS behave very well, no need to speculate beyond that in my view.

I do own a D7000 and can testify that the DR is excellent with very low shadow noise.

Cheers,
Bernard

[Ben Rubinstein]

Bit confused here. Albeit diffraction will rob resolution however my experience has been that when compared like for like, at any given aperture, you will have more resolution with a higher megapixel count than with a lower megapixel count. That your 1Ds3 now only has 18 megapixels doesn't mean that at the same aperture, a 1Dx will retain 18 megapixels without any degradation. Am I wrong in that?

[allegretto]

There are two dominant sources of noise in typical images -- photon noise and read noise.  If we're only concerned with photon noise (ie there are enough photons that read noise is a secondary consideration), then smaller pixels are more 'informative'.  The number of photons captured is the same, they are simply partitioned into samples of finer spatial resolution, so the total captured information is larger due to the better spatial localization of any given captured photon.  Read noise, on the other hand, is a constant noise per pixel, so unless the smaller pixels have smaller read noise in proportion to the pixel spacing, they degrade the image by increasing the noise relative to the number of photons captured.

As usual, there are caveats.  If pixels get too small, there can be cross-talk -- red photons leak into green pixels etc, resulting in a loss of color separation and resolution.  Also, lower S/N per pixel leads to a greater degree of interpolation error in the demosaic process, and thus more 'digital artifacting', though that artifacting is pushed off to finer spatial scales where it may be less noticeable.

Yes, thanks

This is what I was trying to point out but perhaps my words were ambiguous. While "resolution" may be enhanced by smaller pixels, more pixels means more noise if we assume the read noise is the same per pixel. Thus the better high ISO performance of bigger and fewer pixels in a given chip size.

[allegretto]

halving the photosites area will give about half the election capacity per photosite, but will also give twice as many photosites detecting light from the same part of the image, so the "maximum photo-electron count" for any given part of the image is about the same; it is just divided up as the sum or more, smaller numbers, and this retains more information about where the light comes from.

yes, but as Emil points out, more pixels = more read noise

wondering if it is possible that advances in material science will allow more efficient (as in DR and S/N) chips in the future. It seems to me that if a given pixel could trap more photons and/or more efficiently convert photons to electron emission it would enhance characteristics... no?

[BernardLanguillier]

I am really not sure about the value of speaking in theoretical terms while we able to conduct an easy experiment with APS sensors...

Cheers,
Bernard

[BJL]

yes, but as Emil points out, more pixels = more read noise

And as with any trade-off, one has to quantify the opposing advantages and disadvantages, not just declare that any amount of additional noise is more important that amount of additional resolution [or vice versa].

With read noise these days at the level of only a couple of elections, it is vastly outweighed by photon shot noise until you are at light levels many, many stops below mid-tones, in the realm tat is almost always printed totally black. Say a DSLR pixel has a well capacity of 40,000. At full well, shot noise is sqrt(40000) or about 200 electrons. At 8 stops down, more than four stops below mid-tones, which is the traditional black level or noise floor in measuring the ISO speed of film, and in a straight print is totally black, the signal is still about 160 electrons, shot noise of about 14 electrons, still overwhelming read noise.

Only at about 12 stops below maximum, signal about 10 electrons, is read noise significant compared to photon shot noise ... and by then S/N ratio due to read noise alone is a pathetic 3:1, so it is going to be ugly anyway if you attempt to print those deep, deep shadows at anything above pure black.

[ejmartin]

And as with any trade-off, one has to quantify the opposing advantages and disadvantages, not just declare that any amount of additional noise is more important that amount of additional resolution [or vice versa].

Agreed.

With read noise these days at the level of only a couple of electrons, it is vastly outweighed by photon shot noise until you are at light levels many, many stops below mid-tones, in the realm tat is almost always printed totally black. Say a DSLR pixel has a well capacity of 40,000. At full well, shot noise is sqrt(40000) or about 200 electrons. At 8 stops down, more than four stops below mid-tones, which is the traditional black level or noise floor in measuring the ISO speed of film, and in a straight print is totally black, the signal is still about 160 electrons, shot noise of about 14 electrons, still overwhelming read noise.

Only at about 12 stops below maximum, signal about 10 electrons, is read noise significant compared to photon shot noise ... and by then S/N ratio due to read noise alone is a pathetic 3:1, so it is going to be ugly anyway if you attempt to print those deep, deep shadows at anything above pure black.

That really depends on the degree of pattern noise in the read noise.  Being correlated noise, it is not properly measured by the std dev; just look at the 5D2, for which the read noise is around 2.5 electrons, but 8 stops down it's a mess.  I don't think read noise is the main issue any more, it's pattern noise.  As you say, random read noise is well below the level that might cause significant issues.

[BJL]

Agreed.

That really depends on the degree of pattern noise in the read noise.  Being correlated noise, it is not properly measured by the std dev; just look at the 5D2, for which the read noise is around 2.5 electrons, but 8 stops down it's a mess.  I don't think read noise is the main issue any more, it's pattern noise.  As you say, random read noise is well below the level that might cause significant issues.

True: my rough numbers were just to indicate the general idea that with good modern sensor designs (maybe than now means Sony/Nikon more than Canon??), the controllable noise from electronic sources (as opposed to the uncontrollable photon shot noise) is only significant at light levels many stops below full exposure, and so mostly of concern to low-light, high ISO needs, whereas a great proportion of situations that would benefit from more than 18MP can be handled at low ISO speeds.

About pattern noise though: can you comment on the main causes, and how it is likely to scale with pixel size and pixel count? I would not assume that electronic noise is at a fixed level per photo-site regardless of photo-site size, because for one thing, larger photo-sites with larger electron capacities seem likely to need larger circuits and semi-conductor components with greater charge/current/voltage capacity, which could increase their noise levels (measured in electrons). Dark current for example has a roughly "square root of well capacity" scaling law, from spec. sheets I have compared.

[ejmartin]

True: my rough numbers were just to indicate the general idea that with good modern sensor designs (maybe than now means Sony/Nikon more than Canon??), the controllable noise from electronic sources (as opposed to the uncontrollable photon shot noise) is only significant at light levels many stops below full exposure, and so mostly of concern to low-light, high ISO needs, whereas a great proportion of situations that would benefit from more than 18MP can be handled at low ISO speeds.

About pattern noise though: can you comment on the main causes, and how it is likely to scale with pixel size and pixel count? I would not assume that electronic noise is at a fixed level per photo-site regardless of photo-site size, because for one thing, larger photo-sites with larger electron capacities seem likely to need larger circuits and semi-conductor components with greater charge/current/voltage capacity, which could increase their noise levels (measured in electrons). Dark current for example has a roughly "square root of well capacity" scaling law, from spec. sheets I have compared.

I'm really not competent to answer on causes.  Some phenomenological observations: Pattern noise is much worse on the 5D2 than 1Ds3, if I recall correctly the latter has twice as many readout channels (8 vs 4) so the data rates are half.  The Sony Exmor sensors have per-column ADC's, so data rates that are in the kHz range rather than MHz range of a standard Canon design (or D3 type design for Nikon).  So the suspicion is that lower data rates allow better control of fluctuations in the normalization of the readout in each channel.

[torger]

Pattern noise is much worse on the 5D2 than 1Ds3.

And the 7D is even worse than 5D2... concerning pattern noise I don't trust Canon until they show something that works. It seems to me that they focus a lot on software noise reduction and brag about clean JPEGs, but as a RAW shooter using third party software I rather want sensor hardware that is state of the art concerning noise performance. I can manage with both 5D2 and 7D and rarely get real problems in prints, but it seems like a waste when Sony etc shows that low noise levels are indeed possible.

[allegretto]

And as with any trade-off, one has to quantify the opposing advantages and disadvantages, not just declare that any amount of additional noise is more important that amount of additional resolution [or vice versa].

With read noise these days at the level of only a couple of elections, it is vastly outweighed by photon shot noise until you are at light levels many, many stops below mid-tones, in the realm tat is almost always printed totally black. Say a DSLR pixel has a well capacity of 40,000. At full well, shot noise is sqrt(40000) or about 200 electrons. At 8 stops down, more than four stops below mid-tones, which is the traditional black level or noise floor in measuring the ISO speed of film, and in a straight print is totally black, the signal is still about 160 electrons, shot noise of about 14 electrons, still overwhelming read noise.

Only at about 12 stops below maximum, signal about 10 electrons, is read noise significant compared to photon shot noise ... and by then S/N ratio due to read noise alone is a pathetic 3:1, so it is going to be ugly anyway if you attempt to print those deep, deep shadows at anything above pure black.

well you fellows know more than me, but if what you say is true I guess we should not be seeing ANY noise since shot noise is so low and read noise virtually non-existent?

[BJL]

... if what you say is true I guess we should not be seeing ANY noise since shot noise is so low and read noise virtually non-existent?

Shot noise is not always so low as to be invisible, especially as the ISO goes up. With my example of a 40,000e- well capacity, 9 stops below maximum gets down to about 80e- signal, 9e- noise, so S/N ratio about 9:1, which is poor. Not much of a problem at base ISO, where this noise is in features about six stops below midtones, and so will normally be printed totally black. But increase the ISO speed by about 4 stops (say from a base of 200 to ISO 3200) and that 9:1 S/N ratio is in light shadows, only about 2 stops below midtones, so you will probably see it. By ISO 12800, the midtones from those 40,000e- photosites are below 10:1 S/N ratio, and probably do not look very good without trading off some resolution to noise reduction or down-sampling ... or changing to fewer, bigger photosites.

[BJL]

So the suspicion is that lower data rates allow better control of fluctuations in the normalization of the readout in each channel.

I can only hope that this is true. And with Sony, Nikon, and Panasonic adopting this aproach of "amplify and digitize as soon as possible with thoasands of ADCs operating at low clock rates", maybe Canon will follow soon.

[ejmartin]

I must have missed something -- since when is Panasonic using this approach?

[BernardLanguillier]

I can only hope that this is true. And with Sony, Nikon, and Panasonic adopting this aproach of "amplify and digitize as soon as possible with thoasands of ADCs operating at low clock rates", maybe Canon will follow soon.

Could this thing called "patent" make this not obvious?

Cheers,
Bernard

[BJL]

Emil,
    I believe that Panasonic uses this approach in its more video-oriented m4/3 sensors, in the GH2 and such, though this is a bit speculative; all I have read for sure is that the GH2 does ADC on-chip, whereas other m4/3 sensors do it off-board IIRC.

Bernard,
    It seems that patents rarely stop big companies from copying ("stealing" in Steve Jobs' words) important new ideas, either by changing enough details of the implementation, or by hiring enough lawyers. For example, no one had an exclusive on CMOS sensors for very long.

[torger]

There's another aspect to noise too, landscape photographers often significantly dynamically compress what was captured. Back in the analog days these was made using gradient filters to lower the sky 2 - 3 stops. Today compression is made when processing the raw file (although some still use gradient filters too) which puts much tougher requirements on S/N ratio.

[Keith Reeder]

And the 7D is even worse than 5D2... concerning pattern noise

Oh, stop it Torger, that's simply not true.

Your 7D might have a problem, but across the board there are far, far more 5D Mk II users that have reported pattern noise "problems" than have 7D users.